Give uses within a loop a greater weight when doing representation inference.

Give uses within a loop a greater weight when doing representation inference.

Uses of a value are weighted by a factor of FLAG_loop_weight (default: 10) for
every loop they are in. This makes uses in inner loops "more important", which
should improve the result of the representation inference.

Committed: https://code.google.com/p/v8/source/detail?r=10573

[Sven Panne, 2011-10-14 09:11:06]

In theory, this should improve some benchmarks, and have at least no negative
effects on others, but measurements show no real conclusive results. :-(

I just wanted to upload my changes to see if somebody sees a stupid mistake or
some misconception how things work...

[fschneider, 2012-01-27 22:33:24]

lgtm.

Have you tried other values than 10? I'm not sure if we actually need a flag
once we figure out a reasonable value for the weights.

https://chromiumcodereview.appspot.com/8277031/diff/1/src/hydrogen-instructio...
File src/hydrogen-instructions.cc (right):

https://chromiumcodereview.appspot.com/8277031/diff/1/src/hydrogen-instructio...
src/hydrogen-instructions.cc:77: static const int weights[] = { 1, w, w*w,
w*w*w, w*w*w*w };
I think no need for static here. Actually, I'd also be fine with hard-coding the
weight to 10 here, so that the compiler can inline the constants. Once set, we
probably don't need to change it often.

[Sven Panne, 2012-01-30 07:51:04]

Currently Golem is still extremely flaky, so I have a slightly uneasy feeling
about enabling this by default. Nevertheless, I think that accounting for loops
in the representation inference somehow is the way to go. Therefore I propose
the following: Let's change the default weight to 1, effectively disabling
things, and play around with the weights a bit until we are sure what a good
value is. Then we can replace the command line flag by a constant and simplify
the code a bit.

Does that sound reasonable?

https://chromiumcodereview.appspot.com/8277031/diff/1/src/hydrogen-instructio...
File src/hydrogen-instructions.cc (right):

https://chromiumcodereview.appspot.com/8277031/diff/1/src/hydrogen-instructio...
src/hydrogen-instructions.cc:77: static const int weights[] = { 1, w, w*w,
w*w*w, w*w*w*w };
On 2012/01/27 22:33:24, fschneider wrote:
> I think no need for static here. Actually, I'd also be fine with hard-coding
the
> weight to 10 here, so that the compiler can inline the constants. Once set, we
> probably don't need to change it often.

Well, initializing the array every time wouldn't be very cool, either, so I went
for a 'static' (What I really wanted to write is 'iterate (*w) 1' ;-).  But
you're right, when we figured out a sensible weight, we should probably change
this to a simple switch on block()->LoopNestingDepth() and use a real constant.